Negotiators from the European Parliament and the EU Council left at around 1pm on Thursday 7 December – after 22 hours of discussions – without reaching an overall agreement on the legislation on artificial intelligence (‘AI Act’) (see EUROPE 13307/3). While the European co-legislators reached a provisional compromise on a number of points, such as the thorny issue of foundation models (see EUROPE 13306/5), work is still required on several major elements, including national security issues and prohibited practices.
The break in negotiations will be short-lived, however, as the co-legislators will resume discussions on Friday 8 December at 9am in a bid to conclude matters. The MEPs involved will start their day an hour earlier, to take stock of the issues still outstanding.
For the time being, the European Parliament and the EU Council have reportedly reached agreement on the approach to foundation models. This is reportedly ultimately based, broadly speaking, on a tiered approach, as suggested by the European Commission in its attempt to strike a balance between the position of the Spanish Presidency of the EU Council – which wanted stricter rules for high-impact models – and the position of France, Germany and Italy, which wanted to introduce codes of conduct and prevent sanctions from being imposed in the first instance (see EUROPE 13297/24).
In practical terms, all models would be subject to requirements, such as publishing a summary of training data, or ensuring that all AI-generated content is clearly identifiable.
For models representing a higher risk, the text provides for an obligation to assess the model and monitor systemic risks. Provisions have also been introduced for cybersecurity issues and, at the Parliament’s request, for the publication of reports on the energy consumption of the AI models in question.
For these models, their definition would be based on the amount of compute used in training. Codes of conduct could complement the rules, but could not replace them.
The negotiators also addressed the issue of the European AI Office. This would come under the European Commission and could have its own budget line. It would be made up of the national competent authorities, and a group of experts would be tasked with advising on the implementation of the text and warning of the risks.
Striking a balance between prohibited practices
Several questions remained unanswered after the first – and long – phase of this potentially final round of interinstitutional negotiations. The European Parliament and the EU Council will attempt to reach agreement on the prohibited practices at this second meeting aimed at finalise matters.
During the discussions, the EU Council tried to remain as close as possible to its initial position, while the European Parliament reportedly wants the list of prohibited practices to be much longer than the EU27 are prepared to accept.
In particular, the Parliament would like to ban emotional recognition software from the workplace, schools, law enforcement agencies and migration control agencies. The EU Council would be prepared to accept a ban in the first two cases, but is reluctant to agree to a ban in the last two.
Similarly, the European Parliament’s position includes a ban on predictive policing, whereas the EU Council would like to limit this ban to investigations based solely on the predictions of the AI system.
The use of remote biometric identification was also the subject of much debate in the Parliament at the start of the legislative process (see EUROPE 13201/1). The European Parliament is in favour of a total ban, except in cases of serious crime, while the EU Council would like to see greater freedom in the use of this practice.
It remains to be seen whether the points on which the negotiators have agreed will remain unchanged at the end of the forthcoming discussions, as no subject is truly finalised until an agreement has been reached on the text as a whole. (Original version in French by Thomas Mangin)